Research On Deep Embedded Clustering Algorithm Based On Autoencoder

Cluster analysis is a widely used statistical technique.Clustering performance is largely determined by the quality of data representation.Therefore,linear or nonlinear feature transformation has been widely used for better data representation of learning clustering.In recent years,with the increase in information acquisition methods,high-dimensional data has become difficult to cluster.Thanks to the development of deep learning,deep neural networks extract the feature space of original data instead of traditional clustering methods to help better clustering performance.This paper has done an in-depth investigation and research on the deep clustering algorithm,the main work includes:(1)A deep embedded clustering model based on variational autoencoder is proposed.Firstly,the original data features are extracted through the image structure of the variational autoencoder,and then the feature space are clustered through the deep clustering layer to propagate the clustering loss.The deep clustering layer and the variational autoencoder work together to improve the representation of feature space to better learn the original data and the final clustering results.(2)A multi-modal deep embedded clustering model based on autoencoder is proposed.The model performs adaptive feature fusion which is extracted from autoencoder,convolutional autoencoder,and variational autoencoder models.The merged features are passed through the deep clustering layer to obtain the clustering results.Multi-modality can extract different feature information of data,and adaptive feature fusion can well calculate the contribution of each modal feature to the fusion feature,which improves the unsupervised algorithm more stable.The above two models proposed in this paper are tested and compared on four public datasets.The experimental results illustrate the effectiveness of the model proposed in this paper,and the clustering results are better than other popular algorithms.